# Tara selection and Lilith handoff · delivered increment · 8 September 2026

This package continues the [full V1 program](v1-completion.md) after service
search, saved items and launch. Implementation baseline:
`0a6efd8473a7497da0660d04f1953d245eaa2ecd`; primary source pin:
`1ceae52c7896a59ac74a4429d410f3218f0b91c6`.

Audience: product, design and engineering readers who have completed the Tara
catalog and discovery guides. Outcome: understand the complete feature taxonomy
reference, distinguish suggestion metadata from available modalities, and trace
the returned Lilith handoff through its actual selection and rule boundaries.

Visual thesis: a readable paper field guide alternates compact reference tables
with native record specimens, conditional branches, shared-input relationships
and a priority diagram. Use actual source data and declared cases; no invented
product screen or implied clinical efficacy. Theme and modality references need
space to expose their differences, so they belong in a selection guide with a
linked handoff guide rather than one compressed overview.

Content plan: show the three feature models and a slate record; cover all twelve
moods and thirty-seven slate entries; explain effective distress and the
retained slate; compare mood slates with compatible themes; cover all fifteen
themes and fourteen modalities; inspect cadence, constraints, fallback metadata
and validation; close the restless worked case. The handoff guide follows input
normalization, type precedence, utterance/context admission, detector order,
trigger/reason distinctions, rule summaries and all thirteen selectable rule
types; it ends at resource planning and actual application enforcement.

Interaction thesis: keep native relationships readable in PDF and narrow views,
use existing source/explanation reveals and reduced-motion navigation, and link
the guides through prerequisites. New shared renderer behavior is unnecessary.
The existing recurrence treatment represents optional activities, so it is not
appropriate for mandatory cadence phases; use an explicit field specimen for the
data record instead of drawing an unsupported playback cycle.

## Current evidence

The complete mood, theme, modality and Tara–Lilith handoff modules have been
read. Read all thirteen policy rule bodies, the complete signal taxonomy,
detector and matching functions, rule getter and validators. The
regional/response/follow-up, frame and no-bypass helper bodies were read in the
prior delivery and remain recorded in
[source discovery](v1-tara-selection-discovery.md). The complete resource
contact registry is not accepted as current contact guidance.

Five unmodified owning test files run in a temporary pinned source tree through
the installed Vitest runner: 64 tests pass. Explicit feature/persona facades
expose only complete owning modules; unrelated barrels and application
integrations are outside the fixture. Twenty-one additional source cases
exercise five complete modules, including all 144 mood/distress/classifier
combinations, retained slates, constraint exclusions, independent theme
defaults, validation omissions, classifier normalization, context gating, mixed
detector order, mismatched reason/selection and every explicit crisis type. Ten
source/test paths are exact between the pin and baseline. Source cases use
declared data, with no classifier model, device, service/database, navigation or
clinical acceptance.

The scoped pinned symbol search again found the Tara handoff implementation and
feature helpers, but no app invocation in apps/libs for the searched names.
Runtime suppression, actual start controls, persistence, resource availability,
review and recovery need their owning consumers before integration acceptance.
Full Lilith policy and recovery teaching remains part of the wider V1 program.

## Delivery plan

Review complete drafts at desktop and narrow widths before generating audio.
Then promote the linked guides, format canonical JSON, render new narration in
serial batches, and run the full affected browser/center/PDF/audio avenues.
Preserve all 746 earlier slides, scripts and recordings and all 68 earlier guide
HTML files. Retain source, optical and automated evidence, publish with normal
hooks, and keep full V1 acceptance active until the remaining workstreams pass.

## Delivery evidence

The [receipt](../verification/v1-tara-selection-2026-09-08/README.md) binds 25
selection slides and 19 handoff slides, with 3508.925 seconds of new narration.
The source fixture passes 64 unmodified owning tests and 21 additional cases
over five complete modules; ten source/test paths match the pin and baseline.
Final browser review passes 316 checks with 176 captures, 132 embedded axe runs,
eight enlarged-text cases and five rejected missing-content controls. Focused
center review covers both guides, all 44 slides in two embedded contexts, four
audio boundaries and 44 PDF pages. All 68 earlier guide HTML files and four
shared UI files remain exact. The settled presentation suite passes 119 cases.

Optical review of the first drafts identified repeated comparison takeaways.
Each now states the slide’s specific lesson; source captions retain their scope.
The anxious slate uses four actual candidate rows. The priority ladder uses
shorter branch labels and condition descriptions, preserving every branch. The
final rule references also expose every required resource kind on the slide.
Mood fallback now precedes text detection and the combined priority diagram, so
readers inspect each input path before selection. Comparison headings describe
their fields, peer emphasis is null, and only the selected-rule validation
boundary has deliberate priority emphasis. Final PDF and browser captures are
recorded separately from the earlier draft images.

In-app browser interaction was unavailable because the required JavaScript tool
was not exposed. Repository Playwright automation and rendered-image inspection
provided browser and optical review. No human listening/comprehension, live
member application or clinical acceptance is claimed. Full V1 work stays active.
